Assess the concordance between VCFs

library(vcfppR)

Background

In a benchmarking, it’s often that we need to calculate the concordance rate between the test set and truth set. In the truth VCF file, there are always true genotypes GT otherwise we can’t validate the test VCF. A test VCF may be from the variant caller or the genotype imputation program, and the format can be variable, e.g GP, DS and GT. Hence, this article guides you how to use the vcfppR::vcfcomp function to rapidly examine various statistics for different scenarios and formats, such as the Pearson correlation of genotyping (stats=“r2”), the Non-Reference Concordance (stats=“nrc”), the F1-score (stats=“f1”) or the Phasing Switch Error (stats=“pse”).

Case 1: Genotype imputation accuracy

We normally get genotype posterior GP and genotype dosage DS from the diploid imputation software, eg QUILT and GLIMPSE. To examine the imputation accuracy, we calculate the Pearson correlation between the imputed genotype dosage and the true genotypes. With vcfcomp, we need to specify the desired stats="r2" and formats=c("DS","GT"), which will extract the respective FORMAT items for the testvcf and truthvcf.

vcfcomp(testvcf, truthvcf, formats = c("DS", "GT"), stats = "r2")

Besides, the QUILT2-nipt method outputs MDS and FDS for both maternal and fetal genotype dosages in constract to the DS in diploid mode. To assess the imputation accuracy of both the maternal and fetal, we only need to specify the corresponding formats.

vcfcomp(testvcf, truthvcf, formats = c("MDS", "GT"), stats = "r2")
vcfcomp(testvcf, truthvcf, formats = c("FDS", "GT"), stats = "r2")

Case 2: Genotype concordance

In this case, we are interested in the called genotype concordance and the sensitivity / specificity in genotype calling. In addition to stats="r2", we choose stats="f1" or stats="nrc" and specify the formats=c("GT", "GT"). Normally, we want the results for each sample, which can be achieved by using by.sample=TRUE.

vcfcomp(testvcf, truthvcf, formats = c("GT","GT"), stats="nrc", by.sample=TRUE)
vcfcomp(testvcf, truthvcf, formats = c("GT","GT"), stats="f1", by.sample=TRUE)

Case 3: Phasing switch error

In this case, we look for a functionality to assess the phasing switch error(PSE). First of all, we need the two VCF files to contain the phased GT, which is represented through the ‘|’. We can choose to return the sites that have PSE.

vcfcomp(testvcf, truthvcf, stats="pse", return_pse_sites=TRUE)

Note: Currently, the pse statatics is a simple form that doesn’t take the completeness and quality into account.

Case 4: Multiple testing repetitively

In the comprehensive benchmarking, we often run many tests against the same true sets. In this scenario, we can save the truth object and reuse it. Actually, both test and truth can take as input a vcftable object or a RDS file. The RDS file stores an object that returned by vcftable.

saveRDS(vcftable(truthvcf), "truth.rds")
vcfcomp(test=testvcf1, truth="truth.rds")
vcfcomp(test=testvcf2, truth="truth.rds")
vcfcomp(test=testvcf3, truth="truth.rds")

Note: Since the input is R objects instead of VCF path, then you can not use the formats option in vcfcomp to specify new formats other than the ones in your input R objects.

Q&A

Where is the allele frequency from and can I use customized one?

In default, the allele frequency of each vairant is calculated on the fly from the truth VCF based on GT. Yes, you can specify an external file that stores allele frequencies from a database. Currently, vcfcomp can only take a space-separated text file with five columns and a header named with “chr” “pos” “ref” “alt” “af”.

How to fix the error “inconsistent samples name”

If one is certain about the samples name in the testvcf can be replaced with other names that can match the samples in the truthvcf. One can use the names option to specify a vector of new names that can be found in the truthvcf.

NA values returned

  1. make sure you read the docs carefully and use the correct FORMAT in your VCF files
  2. set bins to bigger intervals because there will be NAs if no variants exist in the specified intervals bin.